Kernel interpolation in Sobolev spaces is not consistent in low dimensions

摘要

We consider kernel ridgeless ridge regression with kernels whose associated RKHS is a Sobolev space $H^s$. We show for $d/2